_Inspired by imagenet-multiGPU.torch, uses torchnet dataloaders and engines.
- ClassificationLogger is a combination of meters which can be used anywhere.
- RandomDataset is useful for random sampling.
Pretrained Alexnet using this code
(Logs at http://imgur.com/a/QWlYr)
(Final Train Acc: 59.33% | Val Acc: 58.59% | Val mAP: 0.62 after 55 epochs)
- Note
- For the script to work without GPU, lots of lines (like
cutorch.synchronize()
) will have to be removed. - You will need to prepare the val folder into the flat folder style. See imagenet-multiGPU.torch for instructions.
- setGPU() does not work with multiple GPUs (probably change the createDataParallelTable to setGPU() )